TikTok on its Way to Becoming a Super-App and Google's Founders Flee California

TikTok's Path to a Super-App

TikTok is often described as a social media giant, but the app has long since moved beyond that category. Over the years, it has added TikTok Shop, a map for local discovery, an expanded search function, games, and more. Most recently, hotel bookings and an effort to obtain a fintech license have followed. The model is WeChat in China, a platform that combines Facebook, WhatsApp, Apple Pay, and an App Store. After the major step with TikTok Shop, the company, which has been primarily US-owned since January, has applied the same playbook to new areas. In sports, for example, a dedicated FIFA World Cup hub with scores, schedules, and highlights ran in early June, enabled by the TikTok GamePlan product and partnerships with MLS and MLB. → Techpresso

Synthszr Take: The interesting question isn't whether TikTok can build a super-app, but why no one has managed to do it in a Western market so far. WeChat works in China because it captured the user interface early on, before banks, retailers, and authorities could establish their own apps. In the West, that space is already taken: Apple and Google control the payment and identity layers, and they know their leverage well. TikTok's real moat is its attention stream, where a hotel booking or World Cup score appears incidentally without you ever leaving the app. This is exactly what the toothbrush test describes: a product you use daily that changes your behavior, not just another icon on your homescreen. Whether the fintech license gets approved is the real test, because without a payment function, the super-app remains an ecosystem for videos with a shop bolted on. Whoever owns the touchpoint where consumption and payment merge wins; everything else is featuritis in fast-forward.

Google Founders Flee California's Billionaire Tax

California's voters will decide on the California Billionaire Tax Act in November, a one-time 5 percent levy on anyone worth more than $1.1 billion who has lived in the state since January 1. The initiative, introduced by a healthcare union, is intended to draw about $100 billion from the state's approximately 200 billionaires over five years, with 90 percent going to the healthcare system. It is supported by Ro Khanna, Tom Steyer, and Bernie Sanders. Governor Gavin Newsom, on the other hand, warns of an exodus and prefers to propose a national version (which is unlikely to pass under Trump). Google co-founder Sergey Brin, the third-richest person in the world, has already moved to Nevada and has invested more than $100 million in rival initiatives through a counter-committee. Approval currently stands at 54 percent. In parallel, OpenAI is considering postponing its IPO to 2027: its last valuation was $730 billion, and Altman wants a trillion. → Morning Brew

Synthszr Take: The most interesting number in this story isn't the tax, but Brin's moving van. Wealth held in software and equity is more mobile today than ever, and changing residency to Nevada is cheaper than a double-digit billion-dollar sum. This is precisely what makes a state-level wealth tax so shaky: it taxes people who can become residents elsewhere with two clicks, while their money continues to work in the same Californian AI industry. Newsom has understood this, hence his evasive maneuver toward the federal level, which no one can leave. The real conflict lies elsewhere: California is financing its healthcare system with a handful of fortunes created from AI valuations whose half-life can be seen in Oracle's 19-percent week and OpenAI's postponed IPO. Anyone who builds a one-time tax on paper assets that don't even know if they're worth $730 billion or a trillion is planning on a foundation that reshuffles itself daily. Sustainable state finances come from broad, stable sources, not from a snapshot of the two hundred richest people on a specific date.

Apple Seeks 'Green Light' for Chinese Memory Chips

Apple has reportedly asked the Trump administration for a green light to purchase memory chips from the Chinese manufacturer CXMT, which is on the U.S. trade blacklist. The background is a shortage of DRAM and NAND modules, fueled by the enormous memory appetite of AI data centers, which is driving up prices. Apple needs these components for iPhones and other devices in quantities that no single supplier can comfortably provide. The move would force Apple into direct negotiations with the government, as the blacklist prohibits U.S. companies from trading with listed Chinese firms without a special permit. At the same time, the entire industry is grappling with a memory supercycle as AI training and inference consume available capacity. Whether Washington will approve is uncertain. → Techpresso

Synthszr Take: Apple, the company with the best supply chain in the world, has to knock on a government's door to get enough memory. This shows how much the AI boom has upended the physical layer of the industry. For years, compute was the bottleneck; now it's the humble DRAM module, because every data center for training and inference devours the same modules as Apple's iPhone. This is the Jevons paradox in its purest form: the more efficient and cheaper AI becomes, the more hardware its applications consume. Anyone who thought geopolitics and product strategy could be neatly separated sees the opposite here, as Tim Cook negotiates for memory like a diplomat for tariffs. Memory procurement has become a core strategic competency, and Cupertino is feeling that more acutely than it would like. Anyone planning to build AI products in the coming years should take memory purchasing as seriously as the model itself.

Ford Had to Rehire 350 Engineers Because AI Failed

Ford has publicly admitted that its own AI systems were not delivering the quality the company expected. Charles Poon, VP for Vehicle Hardware Engineering, told reporters that they believed they could simply replace experienced engineers with AI and still build a high-quality product. The problem wasn't a broken model, but that experienced people left before their knowledge could be transferred into the training data; the automated tools then amplified weak inputs instead of catching design flaws. Ford rehired 350 experienced engineers, established a 40-person software QA team, and added over 100,000 AI-supported tests. The result: first place in JD Power's 2026 Initial Quality Study with 152 problems per 100 vehicles, ahead of Nissan and Buick, for the first time in 16 years. The company has cut about 5,300 salaried positions since its employment peak in 2020, part of over 20,000 white-collar jobs eliminated in Detroit. CEO Jim Farley had predicted that AI would 'literally replace half of all office workers in the U.S.,' a statement that his own quality crisis now makes look bad. → Techpresso

Synthszr Take: This is the most expensive lesson on AI Debt anyone has made public this year. Ford sent its institutional knowledge out the door and realized that AI is only as good as the judgment you encode into it beforehand. This is exactly what we wrote in January about the commodification of expertise: AI takes over the syntactic doing and increases the value of the remaining human work, instead of replacing it. Farley's half-and-half prediction was glib because it costs nothing as long as someone else foots the bill. Here, the company paid the bill itself, with 51 recalls affecting over 11 million vehicles in 2026, more than double the next manufacturer. The real question isn't reskilling via a $500 million fund like RAISE US, but which people you can't afford to lose in the first place. Answering that before the wave of layoffs, rather than after the quality collapse, saves you from buying them back at double the price.

AI Slop Floods the Self-Publishing Market

In 2025, Rakuten Kobo rejected a whopping 45 percent of all titles submitted to Kobo Writing Life. CEO Michael Tamblyn attributes over 80 percent of these rejections to books he classifies as obviously AI-generated and of very poor quality, totaling hundreds of thousands of submissions. He first mentioned the figures at the CONTEC conference in Buenos Aires in April, then reiterated them in a Threads post in May. Importantly, the 45 percent represents the share of rejected submissions, not the share of actual machine-written books, and Kobo does not publish the total number of files reviewed or its detection method. Tamblyn himself admits that the AI detection software tested regularly mistakes human text for machine output and vice versa, which is why Kobo makes quality, not origin, the reason for rejection. Amazon's Kindle Direct Publishing took a different path, requiring a declaration of AI-generated content upon upload since September 2023. Kobo prefers rejection over a disclosure form. → Techpresso

Synthszr Take: The most honest number in this report is the one Kobo doesn't mention: how many genuine books end up in the filter to keep the junk out. Tamblyn openly admits that AI detection fails in both directions, which is why Kobo judges based on quality rather than origin. This is the only defensible position as long as detection is the weak link. I find the parallel to code interesting: in 2025, Veracode measured that 45 percent of AI-generated code failed security tests, almost the same rate as the reject pile here. AI amplifies what's already there, and it produces mediocrity faster and in larger quantities than any platform can sort. An author who wants to break through needs the same moat as any good developer: a voice that a machine can't replicate in seconds. The distribution of AI texts is free, but a platform's credibility is not, and Kobo has understood which of these two currencies is scarce.

Google Sees 'Vibe Coding' Coming

This week, Google released the whitepaper 'The New SDLC With Vibe Coding,' co-authored by Addy Osmani, Shubham Saboo, and Sokratis Kartakis. The central thesis: an agent is a model plus a harness, with the split being about 10 percent model to 90 percent harness. The harness includes instructions, rule-files, tools, MCP servers, orchestration logic, and observability. Two numbers make this tangible: on Terminal Bench 2.0, a team moved its coding agent from outside the top 30 into the top 5 using the same model, solely through harness changes. LangChain scored 13.7 points on the same benchmark by only changing the system prompt, tools, and middleware around a fixed model. The paper categorizes context into six types (instructions, knowledge, memory, examples, tools, guardrails) and states: verification is the dividing line between Vibe Coding and real engineering. → Nico Lumma

Synthszr Take: The 10/90 split is the most honest statement in the entire paper. Everyone is staring at the next model, but the real leverage is in the harness—exactly the part you can fix yourself this morning without waiting for a new model generation. When an agent messes up, it's usually due to a missing tool, a loosely written rule, or a context window full of garbage. This aligns with what I call Code Crash Agentic Engineering: whoever vibe-codes is just hoping it works; whoever builds and versions the harness ensures it works. The context economy will be interesting, because you pay for static context with every single call, and that hits the bill directly (we wrote about exploding token costs back in early June). The discipline is shifting from writing code to configuring and validating.

Red-Teaming for Agentic Red Teams

A new arXiv paper (submitted on June 23, 2026, by Dario Pasquini and colleagues) provides the first deep security analysis of the most-used agentic systems for offensive security operations. The result is unsettling: most of these tools share the same design flaws. An active adversary can steal API keys, establish persistence, and completely take over the operator's machine, even if the agent is running in a segregated container. The authors describe a complete cyber kill chain for this, from initial LLM-manipulation to lateral movement and persistence, to bypassing guardrails and breaking out of the sandbox. From their analysis, they derive a more robust architecture and concrete design principles intended to close the disclosed attack paths at a structural level. → Techpresso

Synthszr Take: The very tools used to break into other systems are themselves the weakest link. This has a certain logic, as the industry has spent two years building capability with little focus on hardening; an agentic pentester that exposes its own operator's machine is the expensive receipt for that. This is exactly what's meant by AI Debt: features are shipped in fast-forward, while the security debt quietly lands in the backlog. What makes this paper valuable isn't the warning, but the architecture it provides. Anyone using such tools today should treat them like any other privileged process on the network, with minimal rights and no blind trust in the sandbox. This can be checked this week, not after the next security audit in the fall. An attack tool that you haven't secured yourself isn't an advantage, it's an open door.

Algorithm Monocultures in Recruiting are Discriminatory

A new study, published at FAccT 2026, analyzed a rare dataset: 3 million applicants, 4 million applications, all screened by algorithms from the same vendor. The results show significant racial bias. 14.74 percent of applications from Asian candidates and 25.87 percent from Black candidates land in positions that disadvantage these groups according to U.S. anti-discrimination standards. Even more troubling is the homogeneity of the results: 4 percent of all applicants who apply for ten jobs are rejected by all ten—more often than would be expected by chance. Because the same software deterministically makes the same decision every time, the researchers simulated what applicants would have received for all open positions. The consequence: if you want to be seen by a human at all, you have to cast a wide net. → AI Secret

Synthszr Take: This is the amplifier principle in its coldest form. When a single vendor dominates the screening market, a bias becomes not one mistake among many, but a systematic barrier that a human never gets to lift. A single bad recruiter makes one bad decision; an algorithm monoculture makes the same bad decision four thousand times, deterministically reproducible, without a human ever looking. In January, we reported on how Anthropic is reinventing job interviews and Meta is evaluating talent based on token consumption; the logic is the same everywhere, only now the consequences are becoming measurable. The real risk is not that the machine discriminates—humans do that too—but that it does so in lockstep, eliminating the variance that used to at least provide second chances. Anyone who relies on a single vendor for recruiting is buying that vendor's blind spots as a monopoly. Running several models in parallel and conducting random human spot-checks is not a compliance luxury, but the minimum insurance against the day the lawsuit arrives.

Eric Ries Releases Successor to 'The Lean Startup'

Eric Ries, creator of the Lean Startup methodology, founder of the Long-Term Stock Exchange, and co-founder of Answer.AI, has a new book out: 'Incorruptible' (Authors Equity, May 2026), which went straight to the NYT bestseller list. In a conversation with Gennaro Cuofano of The Business Engineer, he makes a structural argument, not a moral one: every observable competitive advantage at Anthropic, Costco, Patagonia, Novo Nordisk, or Tony's Chocolonely can be traced back to two ingredients—structure and ethos. Ries was one of the people who consulted Anthropic's founders on building the company. His historical finding: shareholder primacy, which today seems like an eternal law of nature, is barely forty years old and was established in the 1980s through a series of Delaware court rulings. Until the late 1970s, U.S. corporations had to declare a specific purpose and a public benefit upon incorporation. His thesis: anyone who builds in governance friction upfront (PBC filing, fiduciary commitments, mission-guardian structure) lowers long-term coordination costs. 'Harder is easier.' → The Business Engineer

Synthszr Take: The most interesting sentence is the one about the two-page filing. A legal structure that everyone considers god-given can be re-architected with a two-page filing. This is precisely what connects Ries's argument with what I write in Code Crash about governance: if you're shipping in hours instead of months, you need harder guardrails because there's no time for human intervention. In the AI age, this becomes a procurement issue. The companies that win the talent competition and don't get hollowed out under pressure have aligned their charter, board, and operating culture with their purpose before the stress test, not after. Anthropic understood this, which is exactly why they asked Ries before the architecture was set. The takeaway for anyone building now: structure outlives the founder. Founder Mode only solves today's control problem and is silent on succession. Investing upfront instead of paying a premium in trust and pricing power every quarter—that's the cheaper path.

Turning the Mac into a Time Machine

RetroMac is a menu bar app for macOS 14+ (Apple Silicon and Intel) that, with a click, applies a CRT glow look, VHS flicker, or complete operating system interfaces over the entire screen without a restart. It includes over 30 live shaders from Sony Trinitron to Game Boy, plus themes for Windows 98, XP, Mac OS 9, Mac OS X, BeOS, and Snow Leopard, including Solitaire and Minesweeper. A single indie developer from Germany, Maik, built it as a weekend hack for himself. It's now used by over 5,000 people, was featured on Product Hunt, and has been covered by ifun.de, Caschys Blog, and YouTube. The model: free, no account, no tracking, 20 shaders for free, and if you want more, you toss over a one-time payment of €8.88. No VC, no subscription, no roadmap meetings, but it does have a webcam shader that pipes the CRT image into Zoom, Meet, or Teams as 'RetroMac.' → Techpresso

Synthszr Take: While the industry discusses AI Debt, the EU AI Act, and billion-dollar training runs, a developer from Germany builds a product that 5,000 people love, with a price of €8.88 and zero external capital. This is the underestimated half of the software world: one person, a clear value proposition, no telemetry apparatus. RetroMac doesn't sell productivity; it sells the warm tube glow of the computers many of us grew up with, and that's a surprisingly viable use case. GenAI is making it radically cheaper to build such niche products; the moat today is taste and speed, not an engineering army. Maik doesn't need a €546,000 tool budget for 200 engineers; he needs a good idea and a weekend. Anyone who thinks there's no room left for solo developers in the hyper-competitive world of big tech should take a closer look at that €8.88.